Single Channel Speech Source Separation Using Hierarchical Deep Neural Networks

Single-channel speech source separation is a well known task for preparing speech signals for some applications like speech recognition and enhancement. In this paper, we introduce a novel design for separating sources with the help of hierarchical deep neural networks and time-frequency masks. The proposed method classifies the mixture signals in three categories based on the mixed genders in the first hierarchy. Thus, three other networks, each for a specific mixture type, use these categorized data for speech separation. Then, an enhancement stage improves the quality of voices considering an improved cost function that reduces the interference of the estimated sources of the previous stage. The demanded data is gathered from TSP corpus and the output of the systems have been evaluated with different metrics such as signal to distortion ratio (SDR), signal to interference ratio (SIR) and Perceptual evaluation of speech quality (PESQ). Comparing with other methods, the proposed architecture works considerably better and the results are outstanding.